The recent success of large language models (LLMs) has been predominantly driven by curating the training dataset composition, scaling of model architectures and dataset sizes and advancements in pretraining objectives, leaving tokenizer influence as a blind spot.Shedding light on this underexplored area, we conduct a comprehensive study on the influence of tokenizer choice on LLM downstream performance by training 24 mono- and multilingual LLMs at a 2.6B parameter scale, ablating different tokenizer algorithms and parameterizations. Our studies highlight that the tokenizer choice can significantly impact the model’s downstream performance and training costs. In particular, we find that the common tokenizer evaluation metrics fertility and parity are not always predictive of model downstream performance, rendering these metrics a questionable proxy for the model’s downstream performance. Furthermore, we show that multilingual tokenizers trained on the five most frequent European languages require vocabulary size increases of factor three in comparison to English. While English-centric tokenizers have been applied to the training of multi-lingual LLMs in the past, we find that this approach results in a severe downstream performance degradation and additional training costs of up to 68%, due to an inefficient tokenization vocabulary.

Whether or not several Creole languages which developed during the early modern period can be considered genetic descendants of European languages has been the subject of intense debate. This is in large part due to the absence of evidence of intermediate forms. This work introduces a new open corpus, the Molyé corpus, which combines stereotypical representations of three kinds of language variation in Europe with early attestations of French-based Creole languages across a period of 400 years. It is intended to facilitate future research on the continuity between contact situations in Europe and Creolophone (former) colonies.

This document describes the submission of the very first version of the Occiglot open-source large language model to the General MT Shared Task of the 9th Conference of Machine Translation (WMT24). Occiglot is an open-source, community-based LLM based on Mistral-7B, which went through language-specific continual pre-training and subsequent instruction tuning, including instructions relevant to machine translation.We examine the automatic metric scores for translating the WMT24 test set and provide a detailed linguistically-motivated analysis.Despite Occiglot performing worse than many of the other system submissions, we observe that it performs better than Mistral7B, which has been based upon, which indicates the positive effect of the language specific continual-pretraining and instruction tuning. We see the submission of this very early version of the model as a motivation to unite community forces and pursue future LLM research on the translation task.

The development of large language models (LLMs) relies heavily on extensive, high-quality datasets. Publicly available datasets focus predominantly on English, leaving other language communities behind. To address this issue, we introduce Community OSCAR, a multilingual dataset initiative designed to address the gap between English and non-English data availability. Through a collective effort, Community OSCAR covers over 150 languages with 45 billion documents, totaling over 345 TiB of data. Initial results indicate that Community OSCAR provides valuable raw data for training LLMs and enhancing the performance of multilingual models. This work aims to contribute to the ongoing advancements in multilingual NLP and to support a more inclusive AI ecosystem by making high-quality, multilingual data more accessible to those working with low-resource languages.

We present and describe two language resources in this paper: CATalog 1.0, the largest text corpus in Catalan to date, and CURATE (Corpus Utility for RAting TExt), a modular, parallelizable pipeline used for processing and scoring documents based on text quality that we have optimised to run in High Performance Cluster (HPC) environments. In the coming sections we describe our data preprocessing pipeline at length; traditional pipelines usually implement a set of binary filters such that a given document is either in or out. In our experience with Catalan, in lower-resource settings it is more practical to instead assign a document a soft score to allow for more flexible decision-making. We describe how the document score is calculated and highlight its interpretability by showing that it is significantly correlated with human judgements as obtained from a comparative judgement experiment. We additionally describe the different subcorpora that make up CATalog 1.0.

The successes of contextual word embeddings learned by training large-scale language models, while remarkable, have mostly occurred for languages where significant amounts of raw texts are available and where annotated data in downstream tasks have a relatively regular spelling. Conversely, it is not yet completely clear if these models are also well suited for lesser-resourced and more irregular languages. We study the case of Old French, which is in the interesting position of having relatively limited amount of available raw text, but enough annotated resources to assess the relevance of contextual word embedding models for downstream NLP tasks. In particular, we use POS-tagging and dependency parsing to evaluate the quality of such models in a large array of configurations, including models trained from scratch from small amounts of raw text and models pre-trained on other languages but fine-tuned on Medieval French data.

anguage models for historical states of language are becoming increasingly important to allow the optimal digitisation and analysis of old textual sources. Because these historical states are at the same time more complex to process and more scarce in the corpora available, this paper presents recent efforts to overcome this difficult situation. These efforts include producing a corpus, creating the model, and evaluating it with an NLP task currently used by scholars in other ongoing projects.

The need for large corpora raw corpora has dramatically increased in recent years with the introduction of transfer learning and semi-supervised learning methods to Natural Language Processing. And while there have been some recent attempts to manually curate the amount of data necessary to train large language models, the main way to obtain this data is still through automatic web crawling. In this paper we take the existing multilingual web corpus OSCAR and its pipeline Ungoliant that extracts and classifies data from Common Crawl at the line level, and propose a set of improvements and automatic annotations in order to produce a new document-oriented version of OSCAR that could prove more suitable to pre-train large generative language models as well as hopefully other applications in Natural Language Processing and Digital Humanities.

En dépit de leur qualité certaine, les ressources et outils disponibles pour l’analyse du français d’Ancien Régime ne sont plus à même de répondre aux enjeux de la recherche en linguistique et en littérature pour cette période. Après avoir précisément défini le cadre chronologique retenu, nous présentons les corpus mis à disposition et les résultats obtenus avec eux pour plusieurs tâches de TAL fondamentales à l’étude de la langue et de la littérature.

With the success of large-scale pre-training and multilingual modeling in Natural Language Processing (NLP), recent years have seen a proliferation of large, Web-mined text datasets covering hundreds of languages. We manually audit the quality of 205 language-specific corpora released with five major public datasets (CCAligned, ParaCrawl, WikiMatrix, OSCAR, mC4). Lower-resource corpora have systematic issues: At least 15 corpora have no usable text, and a significant fraction contains less than 50% sentences of acceptable quality. In addition, many are mislabeled or use nonstandard/ambiguous language codes. We demonstrate that these issues are easy to detect even for non-proficient speakers, and supplement the human audit with automatic analyses. Finally, we recommend techniques to evaluate and improve multilingual corpora and discuss potential risks that come with low-quality data releases.

Named entity recognition has become an increasingly useful tool for digital humanities research, specially when it comes to historical texts. However, historical texts pose a wide range of challenges to both named entity recognition and natural language processing in general that are still difficult to address even with modern neural methods. In this article we focus in named entity recognition for historical French, and in particular for Early Modern French (16th-18th c.), i.e. Ancien Régime French. However, instead of developing a specialised architecture to tackle the particularities of this state of language, we opt for a data-driven approach by developing a new corpus with fine-grained entity annotation, covering three centuries of literature corresponding to the early modern period; we try to annotate as much data as possible producing a corpus that is many times bigger than the most popular NER evaluation corpora for both Contemporary English and French. We then fine-tune existing state-of-the-art architectures for Early Modern and Contemporary French, obtaining results that are on par with those of the current state-of-the-art NER systems for Contemporary English. Both the corpus and the fine-tuned models are released.

The French TreeBank developed at the University Paris 7 is the main source of morphosyntactic and syntactic annotations for French. However, it does not include explicit information related to named entities, which are among the most useful information for several natural language processing tasks and applications. Moreover, no large-scale French corpus with named entity annotations contain referential information, which complement the type and the span of each mention with an indication of the entity it refers to. We have manually annotated the French TreeBank with such information, after an automatic pre-annotation step. We sketch the underlying annotation guidelines and we provide a few figures about the resulting annotations.

We introduce the first treebank for a romanized user-generated content variety of Algerian, a North-African Arabic dialect known for its frequent usage of code-switching. Made of 1500 sentences, fully annotated in morpho-syntax and Universal Dependency syntax, with full translation at both the word and the sentence levels, this treebank is made freely available. It is supplemented with 50k unlabeled sentences collected from Common Crawl and web-crawled data using intensive data-mining techniques. Preliminary experiments demonstrate its usefulness for POS tagging and dependency parsing. We believe that what we present in this paper is useful beyond the low-resource language community. This is the first time that enough unlabeled and annotated data is provided for an emerging user-generated content dialectal language with rich morphology and code switching, making it an challenging test-bed for most recent NLP approaches.

We use the multilingual OSCAR corpus, extracted from Common Crawl via language classification, filtering and cleaning, to train monolingual contextualized word embeddings (ELMo) for five mid-resource languages. We then compare the performance of OSCAR-based and Wikipedia-based ELMo embeddings for these languages on the part-of-speech tagging and parsing tasks. We show that, despite the noise in the Common-Crawl-based OSCAR data, embeddings trained on OSCAR perform much better than monolingual embeddings trained on Wikipedia. They actually equal or improve the current state of the art in tagging and parsing for all five languages. In particular, they also improve over multilingual Wikipedia-based contextual embeddings (multilingual BERT), which almost always constitutes the previous state of the art, thereby showing that the benefit of a larger, more diverse corpus surpasses the cross-lingual benefit of multilingual embedding architectures.

Pretrained language models are now ubiquitous in Natural Language Processing. Despite their success, most available models have either been trained on English data or on the concatenation of data in multiple languages. This makes practical use of such models –in all languages except English– very limited. In this paper, we investigate the feasibility of training monolingual Transformer-based language models for other languages, taking French as an example and evaluating our language models on part-of-speech tagging, dependency parsing, named entity recognition and natural language inference tasks. We show that the use of web crawled data is preferable to the use of Wikipedia data. More surprisingly, we show that a relatively small web crawled dataset (4GB) leads to results that are as good as those obtained using larger datasets (130+GB). Our best performing model CamemBERT reaches or improves the state of the art in all four downstream tasks.

This paper investigates the impact of different types and size of training corpora on language models. By asking the fundamental question of quality versus quantity, we compare four French corpora by pre-training four different ELMos and evaluating them on dependency parsing, POS-tagging and Named Entities Recognition downstream tasks. We present and asses the relevance of a new balanced French corpus, CaBeRnet, that features a representative range of language usage, including a balanced variety of genres (oral transcriptions, newspapers, popular magazines, technical reports, fiction, academic texts), in oral and written styles. We hypothesize that a linguistically representative corpus will allow the language models to be more efficient, and therefore yield better evaluation scores on different evaluation sets and tasks. This paper offers three main contributions: (1) two newly built corpora: (a) CaBeRnet, a French Balanced Reference Corpus and (b) CBT-fr a domain-specific corpus having both oral and written style in youth literature, (2) five versions of ELMo pre-trained on differently built corpora, and (3) a whole array of computational results on downstream tasks that deepen our understanding of the effects of corpus balance and register in NLP evaluation.

Les modèles de langue neuronaux contextuels sont désormais omniprésents en traitement automatique des langues. Jusqu’à récemment, la plupart des modèles disponibles ont été entraînés soit sur des données en anglais, soit sur la concaténation de données dans plusieurs langues. L’utilisation pratique de ces modèles — dans toutes les langues sauf l’anglais — était donc limitée. La sortie récente de plusieurs modèles monolingues fondés sur BERT (Devlin et al., 2019), notamment pour le français, a démontré l’intérêt de ces modèles en améliorant l’état de l’art pour toutes les tâches évaluées. Dans cet article, à partir d’expériences menées sur CamemBERT (Martin et al., 2019), nous montrons que l’utilisation de données à haute variabilité est préférable à des données plus uniformes. De façon plus surprenante, nous montrons que l’utilisation d’un ensemble relativement petit de données issues du web (4Go) donne des résultats aussi bons que ceux obtenus à partir d’ensembles de données plus grands de deux ordres de grandeurs (138Go).